ENH: Add annotations for comparison operations #17778
Conversation
| @overload | ||
| def __call__(self, __other: _T) -> bool_: ... | ||
| @overload | ||
| def __call__(self, __other: ArrayLike) -> Union[ndarray, bool_]: ... |
There was a problem hiding this comment.
The comparison ops of generic are one the few that accept actual array-like object (rather than just arrays and/or scalar). Consequently, the signatures are still a bit too broad whenever array-like objects are involved (something to revisit in a future PR after #17719):
import numpy as np
U = np.str_()
i8 = np.int64()
i8 > U # This is considered acceptable (for now) due to the broad definition of `ArrayLike`
| @@ -312,6 +314,12 @@ def __call__( | |||
| class _NumberOp(Protocol): | |||
| def __call__(self, __other: _NumberLike) -> number: ... | |||
|
|
|||
| class _ComparisonOp(Protocol[_T]): | |||
| @overload | |||
| def __call__(self, __other: _T) -> bool_: ... | |||
There was a problem hiding this comment.
Are these overloads in the right order? I can't find docs so quickly, but I had expected overloads to be checked in order, so _T would also match ArrayLike.
There was a problem hiding this comment.
Yes, in practice the first overload is used for all valid scalar-types, which always return bool_, while the second one is used for the much broader ArrayLike (a superset of the first overload).
There was a problem hiding this comment.
Also note that _T is a typevar, so it should be substituted with an actual value.
For example, _ComparisonOp[datetime64] replaces _T with datetime64.
There was a problem hiding this comment.
Yes, in practice the first overload is used for all valid scalar-types
Thanks. That's what I was (am) fuzzy about. The Python docs on TypeVar are pretty bad, and PEP 484 says "By default, a type variable ranges over all possible types.". My interpretation was really "all", not "all scalar". You can use TypeVar( ...bound= with classes, mappings, etc. as bounds. So I thought _T would match array-likes as well here.
There was a problem hiding this comment.
For an unbound (i.e. unassigned) typevars this statement is actually true, even if the bound parameter is specified (because reasons...). This is the very reason why it is recommended to always assign some value to subscriptable types, e.g. use Sequence[Any] instead of plain Sequence.
The above-mentioned fortunately becomes irrelevant once a value has been assigned to the typevar (Seq3 in the example below).
from typing import TypeVar, Sequence
T = TypeVar("T", bound=int)
Seq1 = Sequence # equivalent to `Sequence[Any]`
Seq2 = Sequence[T] # equivalent to `Sequence[Any]`
Seq3 = Seq2[int] # equivalent to `Sequence[int]`There was a problem hiding this comment.
Thanks, that makes more sense now.
|
In it goes, thanks @BvB93 |
This pr adds annotations for the comparison operations of
ndarray/generic:__gt____ge____lt____le__The equivalency operators where deliberately excluded as it is currently impossible to properly type
them due to the existence of
object.__eq__. See #17368 (comment) for more details.